← Back to team overview

dolfin team mailing list archive

Re: Slow PetscInitialize?

 

On Mon, Oct 15, 2012 at 08:35:20PM +0100, Garth N. Wells wrote:
> On Mon, Oct 15, 2012 at 7:59 PM, Anders Logg <logg@xxxxxxxxx> wrote:
> > On Mon, Oct 15, 2012 at 05:28:50PM +0100, Garth N. Wells wrote:
> >
> >> > Do we have that documented somewhere? It would be good to get this
> >> > into Dorsal (if possible) to make it easy to share the knowledge of
> >> > how to build 'hpc' versions of FEniCS on some standard systems.
> >> >
> >>
> >> Not documented, but the issue is simple: with dynamically linked
> >> libraries in standard paths it's mayhem trying figure out what's being
> >> linked when different libraries link to different library versions  of
> >> the same package..
> >
> > Could you share some pointers on how you have set up your build, which
> > versions of packages (MPICH or OpenMPI), which BLAS, version numbers
> > etc and I can try to replicate and think about whether this can be
> > documented as part of Dorsal or somewhere else.
> >
>
> - Most important is ATLAS for BLAS (if MKL is not available)
>
>     http://math-atlas.sourceforge.net/
>
> I use v3.10. It can take a long time to build, and can be tricky (e.g.
> it won't build if the CPU clock frequency is being dynamically
> controlled).
>
> - Build LAPACK to use ATLAS BLAS
>
> - Build latest OpenMPI with support for threads turned on. MPICH2
> works fine too.
>
> - Build Boost so that Boost.MPI uses the manually installed MPI
>
> - Install CGAL to use the manually installed Boost
>
> - Build latest METIS and ParMETIS (Ubuntu ParMETIS is broken, and it
> pull in MPI)
>
> - Build VTK (so that the Ubuntu version doesn't pull in MPI)
>
> - Build HDF5 with parallel suppport  (so that the Ubuntu version
> doesn't pull in MPI)
>
> - Build SCOTCH and PaStiX with support for 64 bit integers
>
> - Build PETSc, SLEPc and Trilinos, linking to the above libraries.
>
> There are probably some tricks to getting around some changes by using
> the Ubuntu 'alternatives' system.

Thanks! I'll try this and report how it goes.

--
Anders


Follow ups

References